Liver Segmentation Using Active Learning Ankur Bakshi Allison Petrosino Advisor: Dr. Jacob Furst August 21, 2008
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Introduction Liver has many important functions Liver cancer is 4 th most common malignancy in the world Computed Tomography (CT) scans are a common tool for diagnosis
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Problem Statement Liver Segmentation is an important first step for Computer-Aided Diagnosis (CAD) Difficulties associated with liver segmentation Time consuming Similarities to other organs Source: Comparison and Evaluation of Methods for Liver Segmentation from CT datasets, Heimann et al., 2008
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Related Work Heimann et al.- statistical shape based segmentation Susomboon et al.- hybrid liver segmentation Tur et al.- natural language application Tong et al.- text classification Turtinen et al.- texture application Prasad et al.- emphysema classification
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Liver Segmentation Algorithm
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Methods Explored Passive Learning Active Learning 1000 vs 100 initial examples 100 vs 10 examples added Negatives taken from evaluated non-liver vs. all non-liver Most informative vs Hierarchical Gabor
Hierarchical Method
Post-Processing
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Results, Patient 1 MethodScore Passive Learning55 Confidence Interval36 Active Learning, 1000 initial examples 81 Active Learning, 100 initial examples 79 Active Learning, 100 examples added 79 Active Learning, 10 examples added 55
Results, Patient 1 MethodScores 10 added non-evaluated added non-liver evaluated 82 Most Informative78 Hierarchical77 Average Human, non- radiologist 75
Results, Patient 1 Slice 134Slice 135Slice 136 Slice 137Slice 138Slice 139
Results, Patient 3 ApproachScores Passive0 Confidence Interval0 Active Learning22
Results, Patient 20 ApproachScores Passive- Confidence Interval59 Active Learning50
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions
Conclusion Classifier based approach outperforms confidence interval based approach Active learning outperforms passive learning Different active learning methods have similar results 10 examples, evaluated non-liver is most promising Interesting structures highlighted for application in CADx systems
Agenda Introduction Problem Statement Related Work Liver Segmentation Methods Results Conclusion Questions